Combining Global and Local Feature Analyses for Quality Evaluation of Tone-Mapped HDR Images

In existing literature, many tone mapping operators have been designed to convert high dynamic range (HDR) images to low dynamic range (LDR) images for visualization on the standard LDR displays. However, it inevitably causes distortions and artifacts due to the dynamic range compression, resulting in an unpleasant visual experience. In this paper, we propose a no reference (NR) quality evaluation method for tone-mapped (TM) HDR images based on global and local feature analyses. The key strategy of the proposed method lies in measuring the abnormal exposure both globally and locally as well as the halo effects. To be specific, given that the abnormal exposure usually induces the color and brightness changes, the global colorfulness and distribution of brightness are first extracted to quantify the abnormal degree of exposure. Then, since abnormal exposure has diverse influences on local regions, the local features are further extracted from the statistical perspective on the discrete cosine transform domain. Finally, the edge strength is computed locally to measure the halo effects. All extracted quality-sensitive features are combined and trained together with subjective ratings to form a regression model using support vector regression. With extensive experiments, we have shown that the proposed method outmatches several mainstream NR image quality assessment methods designed for both natural scene images and TM HDR images.